The seam starts closing

SLOP8217 · Week 5

Four weeks of assumption

A bad translation looks bad.

  • week 2 — a wrong sense, printed
  • week 3 — a rule applied where it did not belong
  • week 4 — a sentence that turns at the join

Every reconstruction so far started from something visibly broken.

2016

What changed

Not accuracy first. Fluency first.

The output became well-formed before it became right, and those turn out to be separate properties.

Two axes, not one

Two axes: reads badly to reads well across the bottom, is wrong to is right up the side. Top-left, is right and reads badly: week 2 to 4 successes. Top-right, is right and reads well: the goal. Bottom-left, is wrong and reads badly: week 2 to 4 evidence. Bottom-right, is wrong and reads well, filled in: this week.week 2–4successesthe goalweek 2–4evidencethis weekreads badly → reads wellis wrong → is right

The bottom-right cell is the one nobody had to think about before.

An error that does not announce itself

A statistical system, starved of input, returns something broken.

A neural system, starved of input, returns a sentence.

It has no other mode. Sentences are what it is for.

Given nothing, it produced scripture.

2018

People fed repeated nonsense syllables into low-resource language pairs.

What came back was fluent apocalyptic prose.

The explanation is not mysterious. For several of those languages, the largest available parallel text was scripture.

I am not in the office at the moment. Send any work to be translated.

The Swansea road sign primary

The output was a portrait of the training set

This is the most legible moment in the whole course.

Nobody published the corpus. The model described it anyway, under conditions where it had nothing else to say.

What that means for the method

The reconstruction still works — but the evidence has moved.

  • before — the error tells you about the table
  • now — the shape of the invention tells you about the distribution

You are no longer reading a mistake. You are reading a default.

And what it costs

The error was a free, involuntary disclosure.

It arrived on every sign, in public, without anybody’s permission.

Week 11 is about what happens when it stops arriving.

At the bench on Wednesday

  1. pick a pair with little parallel text
  2. feed it input with no content
  3. log the system, the direction and the date
  4. repeat one input four times

Step 3 is not admin. In six months your log is the only evidence this behaviour existed.

At the bench on Wednesday

Starve it

Wednesday, 24 March

Material
input with no content
Operation
probing a closed system
Leaves
a dated log of what a system invents